Quaternion Fractional Moment (QFM) descriptors have been widely used for geometric pattern recognition because they can represent multi-channel data and have some degree of invariance. However, they have not been systematically tested for their robustness. It has been particularly unclear to what degree they are robust when the geometric pattern is corrupted by noise influenced by lighting variations, which are common in imaging modalities. A robustness analysis framework for QFM descriptors has been developed in photometric noise-controlled environments. The approach proposes to include bounded spatiotemporal perturbations in the feature extraction process to obtain several realizations of the descriptor for a single geometric pattern. A stability measure indicative of the degree to which a particular feature tends to spread out from the mean of the descriptor, in the context of feature dispersion, ensures that robustness is measured directly in feature space. The classification accuracy is compared to a deterministic QFM and the stability behaviors of the two are linked to that of the QFM. Numerous experiments are conducted on Moroccan-Andalusian geometric patterns with p4m and p6m symmetries. It is supplemented with additive Gaussian noise of varying intensities for photometric stability. As inferred from the results, the proposed approach is able to improve stability in the descriptors with noise and ensure stable performance in recognition accuracy.